Approach A

Approach A - PDDL / Classical Planning

Motivation

The motivation was to develop an agent(s) that uses classical planning, by making calls to a solver using templated environment files, to act within the Pacman contest environment.

Theory

Classical Planning, in the form of model-based planning, solves the problem of which action an agent should do by:

• Specifying a model representing the problem (the state model), including:
  • a finite and discrete state space;
  • a known initial state;
  • a set of goal states;
  • actions applicable in each state;
  • a deterministic transition function;
  • positive action costs;
• Making a call to a planner (using the state model), which returns a controller (i.e. a plan);
• Using this controller to act in the desired environment.

This approach would utilise Planning Domain Definition Language (PDDL) to model the Pacman environment.

Trade-offs

Advantages:

• Requires much less programming (in the form of problem description), as opposed to many other AI methods.
• Can be extremely powerful in some circumstances.

Disadvantages:

• Assumes a single-agent, fully-observable, deterministic, static environment.

Special Note:

This specific approach was the only one which was not successfully implemented in any of the agents; this was primarily due to technical issues experienced in implementing the Metric-FF solver.

Back to AI Approaches Overview